OUTLIERS AND TREND DETECTION TESTS IN RAINFALL

EXTREMES.
PANAGIOTA GALIATSATOU1, PANAYOTIS PRINOS2
1

Ph.D Student, Hydraulics Lab., Dept. of Civil Eng. Aristotle University of Thessaloniki, Greece
(Tel: +30-2310-995856, e-mail: pgaliats@civil.auth.gr)
2
Professor, Hydraulics Lab., Dept. of Civil Eng. Aristotle University of Thessaloniki, Greece
(Tel: +30-2310-995689, Fax: +30-2310-995672, e-mail: prinosp@civil.auth.gr)

ABSTRACT
In the present work daily rainfall time series is analyzed using Extreme Value methodologies.
Four different tests are used for detection of a possible outlier: the estimation of xult, the
conditional probability of the event estimated through a Bayesian formula, a Gumbel plot and
a q-q plot for an appropriately selected model for extreme events. Return level plots are
created using a point process model for extreme observations, both by including and
excluding the possible outlier and discrepancies are discussed. Simple parametric models for
extremal trends are then incorporated in the parameters of the point process model. The
location and scale parameter of the model are assumed to vary as polynomial functions of
time or as sinusoidal terms, treated as entirely distinct or closely inter-related. Alternatively,
they are expressed in terms of two other parameters, the rate and the severity of extreme
events. The deviance statistic is used to identify the significance of such trends. Daily rainfall
from the city of Thessaloniki, Greece is used in the analysis.
Keywords: possible outlier, ultimate event, Gumbel plot, q-q plot, posterior predictive
distribution, return level, non-stationarity, point process, polynomial temporal trends,
deviance statistic
1. INTRODUCTION
Extreme Value Theory is concerned with probabilistic and statistical subjects related to
very high or very low values in sequences of random variables and in stochastic processes.
Environment, finance, Internet traffic, reliability, athletics, statistical inference/tests and
asymptotic theory for sums are some of the main areas of application of Extreme Value
Theory, which is a really rapidly developing subject with a rich mathematical background.
The entire range of possible limit distributions for extreme realizations of a random
variable is given by the Extremal Types Theorem. The families of Extreme Value and
Generalized Pareto distributions are more widely used to perform Extreme Value analyses.
Another particularly appropriate formulation of characterizing the behavior of extremes of a
process is derived from the theory of point processes. Coles (2001) concentrates on statistical
inference for extremes and uses all the previously mentioned methodologies to analyze
different data sets and to produce statistical inference for extremes.
The problem of outliers is of major concern when dealing with extreme events. In
statistics, an outlier is a single observation far away from the rest of the data that can lead to
unrealistic conclusions, especially when considering extrapolation to high enough quantiles of
the variables analyzed. The problem of outliers was considered among others by Smith
(2003), who examined environmental datasets from this point of view, as well as a remarkable
series of track performances achieved by a Chinese athlete including a new world record,
which brought up immediate suspicions. Robinson and Tawn (1995) also examined the
previously mentioned series to assess how much of an outlier this performance was.

In the context of environmental processes, non-stationarity is often apparent because of

seasonal effects, perhaps due to different climate patterns in different months, or in the form
of trends, possibly due to long-term trends (Coles, 2001). The purpose of a trend analysis/ test
is to determine if the values of a series have a general increase or decrease with increasing
time. Trend is often incorporated in the analyses using covariate information. Covariate
modeling in the context of the threshold excess model was discussed by Davison and Smith
(1990). Coles and Tawn (1990) exploit covariate information consisting of location, inter-site
distances and an approximation to the tidal sequence at each site to examine risk assessment
for UK coastlines and the issue of sensitivity to climate change. Dixon and Tawn (1999)
assess the effect of temporal non-stationarity due to a combination of long- term trend in the
mean sea level, the deterministic tidal component, surge seasonality and interactions between
tide and surge. Smith (2003) uses different formulations for the parameters of extreme
models. The covariates here are taken to be polynomial functions of time and sinusoidal
terms. Gaetan and Grigoletto (2004) use a semiparametric approach for smoothing sample
extremes, based on nonlinear dynamic modeling of the Generalised Extreme Value
distribution. Huerta and Sanso (2005) assume that the ozone observations in Mexico city
follow a GEV distribution for which the location, scale or shape parameters define the spacetime structure, where the temporal component is determined through a Dynamic Linear
Model.
In the present study daily rainfall from the city of Thessaloniki, Greece for a time period
of 43 years (1958-2000) is analyzed using Extreme Value methodologies. Four different tests
are used for detection of a possible outlier: a) the estimation of xult, b) the conditional
probability of the event estimated through a Bayesian formula, c) a Gumbel plot and d) a q-q
plot for an appropriately selected model for extreme events of the rainfall process. Return
level plots are presented using a point process model both by including and excluding the
largest observation considered above and discrepancies are discussed. Simple parametric
models for extremal trends are then incorporated in the point process model. These nonstationarities are defined through deterministic functions in the parameters of the model. The
location and scale parameter of the model are assumed to vary as polynomial functions of
time or as sinusoidal terms, treated as entirely distinct or closely inter-related. Alternatively,
they are expressed in terms of two other parameters, the rate and the severity of extreme
events. A statistical parameter, the deviance statistic D, is used to identify the significance of
such trends.
2. POSSIBLE OUTLIERS
2.1. DETECTION TESTS OF A POSSIBLE OUTLIER
Daily rainfall in the city of Thessaloniki is analyzed for a time period of 43 years (19582000). The data are presented in Figure 1 together with the index of day and date, the largest
event occurs.
The event of 23/11/1985, measured at 98mm seems to be significantly larger compared to
the rest of the time series and also to the adjoining events of 22/11/1985 and 24/11/1985. This
is the reason of the following analysis, which is supposed to give some evidence of whether
the event should be included in the analysis or considered to be an outlier and omitted from
the sample.

Figure 1. The rainfall time series of Thessaloniki (1958-2000) and its largest events
An outlier is an observation that lies an abnormal distance from other values in a random
sample from a population. Their inclusion in the analysis can produce deceivable results that
can lead to uneconomical and inappropriate solutions. In the following Sections four different
methodologies are used to produce evidence of whether the event of 23/11/1985 should be
considered an outlier or not: a) the estimation of the nominal ultimate rainfall event, xult b)
the calculation of the conditional probability of the event occurring, through a Bayesian
formula, c) a Gumbel plot and d) a q-q plot for the GPD (Generalized Pareto Distribution) for
the excesses of an appropriately defined threshold of the rainfall process.
2.1.1. THE ULTIMATE RAINFALL EVEN
The data from 1958 up to 1981 are analyzed separately, using a Generalized Pareto
Distribution (GPD) to produce an estimate of xult (median and 95% confidence intervals).
This period is chosen because of the fact that there are no observations that are completely
different from the majority in it. If Yi=Xi-u are the excesses of the rainfall variable X over a
high enough threshold u, in some asymptotic sense, the conditional distribution of excesses
follows the GPD:

y 1
(1)
)

where , denote the scale and shape parameters, respectively (Pandey et al., 2001). Using
the mean residual life plot of the data and the diagram of estimates of parameters and with
different thresholds u, an appropriate threshold uo=16mm is chosen for the short rainfall time
series. By Maximum Likelihood Estimation (MLE) the parameters and their standard errors
)
)
are: =11.931 (se=1.3436) and =-0.1331 (se=0.08016). The shape parameter is dominant
in determining the qualitative behavior of the GPD. If <0 the distribution of excesses has an
upper bound of u-/=105.65mm (Coles, 2001). A profile likelihood for xult may be
constructed, as shown in Figure 2, to reveal a 95% confidence interval for the ultimate return
level.
G(y) = 1 (1 +

Figure 2. Profile log-likelihood for xult

Figure 2 shows a large 95% confidence interval in the range of (60, 260)mm for xult
=105.65mm. The record of 98mm of 23/11/1985 lies within the confidence interval, so on the
basis of the analysis, there is no real evidence that the record is unrealistic. The alternative
Bayesian analysis considers the problem, not as one about estimating the ultimate limit
parameter, but estimates a prediction interval, which will give more precise information about
what is likely to happen in a given year, considering the previously recorded observations
(Smith, 2003).
2.1.2. THE CONDITIONAL PROBABILITY OF THE EVENT TO OCCUR
A primary objective of an extreme value analysis is often prediction. Let z denote a future
observation with density function f(z ) , where . The posterior predictive density of z,
given observed data x is:
f( z x) = f( z )( x)d

(2)

Predictive distributions reflect the uncertainty in the model and the uncertainty due to the
variability of future observations (Stephenson and Tawn, 2004). Using expression (2), the
posterior predictive distribution of a future observation z is given by:
(3)
Pr(Z z x) = Pr(Z z )( x)d

From equation (3) it is obvious that the conditional probability can be expressed as an analytic
function of the three parameters =(, , ) of a point process model, say (,,), and hence
estimated through the Bayesian formula (Smith, 2003):

(, , )(, , x)ddd

(4)

where (... x) denotes the posterior density given past data X. The prior distribution for the
parameters used to construct this posterior density is a trivariate normal distribution. This
distribution enables the specification of independent parameters. Setting =log, we might
choose a prior density function f(,,)=f().f().f(), where f(.), f(.), f(.) are normal
density functions with mean zero and variances , , , respectively, corresponding to a
specification of prior independence in the parameters , and . The absence of genuine prior
information leads to the use of priors that have very high variance, or equivalently, are nearly
flat. The choice of normal densities is arbitrary and variances are chosen as ==104 and
=102 (Coles, 2001).

If ZL is the maximum daily rainfall over a period of L years, then allowing for the
uncertainty in the estimation of the parameter components, the predictive distribution of ZL is
defined as (Coles and Tawn, 1996):
(5)
Pr( Z < z x) = Pr( Z < z ) L ( x)d
L

where Z denotes the annual maximum. If L=1, equation (5) reduces to (2). The result in this
case for P( Z > 98 x) is equal to 0.001111111 for a future period of 1 year. For a future period
of four years P( Z > 98 x) =0.004545455. Both probabilities are not low enough to provide
strong evidence that the event of 23/11/1985 is an outlier.
2.1.3. THE GUMBEL PLOT
Suppose that the annual maxima over the 43 years are Y1, Y2,, Y43 ordered as Y1:43
Y43:43, then Yi:43 for 1 i 43, is plotted against the reduced value xi:43 , where
x i:43 = log(logp i:43 ) , with pi:43=(i-0.5)/43 being the ith plotting position (Figure 3).

Figure 3. Gumbel plot for rainfall data

A straight line indicates a good fit to the Gumbel distribution. There is no systematic
evidence of curvature upwards or downwards in the annual maxima of the rainfall process,
nor do there seem to be evidence of outliers.
2.1.4. Q-Q PLOTS FOR EXCESSES
)
Let F(x) be an estimate of F(x) based on x1, x2,.., xn.The scatter plot of the points
)
F 1 (p i:N ) versus xi:N, i=1,,N is called a q-q plot. Thus, the q-q plot shows the estimated
versus the observed quantiles. If the model fits the data well, the pattern of points on the q-q
plot will exhibit a 45-degree straight line (Castillo et al, 2005). For the GPD model and
)
assuming 0, the quantile plot consists of the pairs: { (H 1 (i/(k + 1)), y(i) } for i=1,.., k
(Coles, 2001), where k is the number of excesses of an appropriately defined threshold. For
two different thresholds of 18mm and 9mm, the q-q plots of points exceeding both thresholds
are shown in Figure 4(a,b).

b. Threshold of 9mm

60

80

20

20

40

Empirical

60
40

Empirical

80

100

100

a. Threshold of 18mm

20

30

40

50

60

70

20

Model

40

60

80

Model

Figure 4. Q-Q plots using two different thresholds

The shape of the plot for both thresholds (especially for the lower threshold) can not
support the treatment of the final data point as an outlier. In fact, the largest point is not very
far from the straight line, especially when considering the lower threshold. It should be noted
here that in the following a threshold of 18mm is used, because it is considered to be more
consistent with the methodology of threshold selection mentioned in 2.1 and described in
detail by Coles (2001). Smith (2003) emphasizes the fact that observations which first appear
to be outliers, may not in fact be inconsistent with the rest of the data if they come from a
long-tailed distribution. By maximum likelihood fitting of the GPD model using both
previously mentioned thresholds, estimates of the shape parameter are 0.0213 and 0.101 for
thresholds 18 and 9mm respectively, confirming a long-tailed distribution.
2.2. RETURN LEVEL ESTIMATION WITH AND WITHOUT THE LARGEST EVENT
The R-year return level is calculated using the formula:
1
)+u
x R = G 1 (1
(6)
R
where G-1(p) denotes the Pareto quantile function. If m is the number of years of observations
and k the total number of exceedances, then the mean crossing rate is =k/m. Figure 5(a, b)
shows estimates of the median and 95% confidence intervals for the return level, when the
largest observation is incorporated or omitted from the sample.
a. Including the observation of 98mm

b. Omitting the observation of 98mm

Figure 5. Return level of rainfall (mm) against return period (years)

Estimates of return levels are generally lower, when the largest observation of 98mm is
omitted from the sample. From Figure 5b, it is easily noticed that the distribution of excesses
of the threshold is now upper- bounded (the shape parameter is negative) in contrast to
Figure 5a, where the observation of 98mm is included in the analysis. 95% confidence bands
in Figure 5a seem not to include the observation of 98mm. These confidence intervals are
estimated assuming a normal distribution for the parameters of the distribution. This is not
considered as a cause of concern, because it is observed that, when using for example the
Bayesian approach with flat priors for the parameters, the most extreme observation is
included in the confidence bands (Figure 6).
Figure 6 shows that the confidence interval constructed using the Bayesian approach
incorporates the largest observation of the 43 years of data available. It can be concluded that
in this case the results of a Bayesian analysis are closer to real observations and to real
conditions, compared to MLE procedure and therefore they lead to a more conservative
solution (confidence intervals include higher values of the rainfall process).
According to Fisher, (1922) the rejection of observations is too crude to be defended and
unless there are other reasons for rejection than the mere divergences from the majority, it
would be more philosophical to accept these extremes, not as gross errors but as the
indications that the distribution of errors is not normal (Reiss and Thomas, 2001). In the rest
of the analysis the observation will not be omitted from the sample, because it is not
considered as an outlier.
ML
Bayes

Figure 6. Return Level estimation (median and 95% confidence intervals) using ML and
Bayes approach
3. DETECTION TESTS FOR TRENDS
Estimates of return levels based on the models analyzed in Section 2 assume that the
behavior of the process will remain unchanged in future years. Many hydrological time series
exhibit trending behavior or non-stationarity. In fact, the trending behavior is a type of nonstationarity. The purpose of a trend analysis is to determine if the values of a series have a
general increase or decrease with time.
If X1, X2,.., Xn is a series of independent random variables with a common distribution
function F, for any large fixed threshold u the sequence {X1, X2,.., Xn}, viewed on the interval
[u, ), is approximately a non-homogeneous Poisson process with intensity function:
(x) =

{1 + (

x
( +1)/
, where >0 and =(, , ) is the vector of parameters
)} +

determined by the tail behavior of F. The approximation outlined leads to a likelihood for
based on an observed set of exceedances x=(x1,x2,..,xnu) of a high threshold u, given by:

n
L( ; x ) = exp( [u, ))
ny

where

is

ny

nu

(x i )

(7)

i =1

the

number

of
observations
per
year
and
and Tawn, 2004). Simple polynomial models

u 1/
(Stephenson
[u, ) = (x)dx ={1 + (
)}+
u

are implemented to describe the structure of (t)=((t), (t), (t)), in order to allow the
parameters of the point process model to vary with time. We are interested to draw inference
about the latent trend (t), where t [t1,..,tn] which corresponds to years 1 to n. Three different
approaches of modeling trends are examined here: a) the modeling of and as polynomial
functions of time when these parameters are treated as entirely distinct, b) the modeling of
and as polynomial functions of time when they are treated as closely inter-related and c) the
modeling of parameters and , which represent the rate at which extreme events occur and
their severity, respectively (Butler, 2005).
3.1. TRENDS IN RAINFALL EXTREMES
3.1.1 PARAMETERS AS DISTINCT POLYNOMIAL FUNCTIONS OF TIME
Model parameters and are treated here as distinct polynomial functions of time of
order p and p respectively, while is supposed to be constant over time (Butler, 2005):
p

k =0

k =0

(t) = (t t 1 ) k k , (t) = (t t 1 ) k k and

(t)=

(8)

This model contains p+p+3 parameters. In the analyses of this paper values of k=1,2 are
implemented to the parameters.
3.1.2. PARAMETERS AS CLOSELY INTER-RELATED POLYNOMIAL FUNCTIONS OF
TIME
This model assumes that (t) is a polynomial of order p and the ratio (t)= (t)/(t) is
constant over time. It follows that (Butler, 2005):
p

(t) = (t t 1 ) k , (t) = (1 / )
k

k =0

(t t )
1

k and

(t)=

(9)

k =0

where is an unknown parameter. This model contains less parameters, namely p+3. The
parameter is called the coefficient of extremal variation and in the following analysis it is
assumed to be constant over time for k=1.
3.1.3. TRENDS IN THE NUMBER OF EXTREME STORMS AND THE SEVERITY OF
STORMS
A widely used model (Coles and Tawn, 1990) treats the rate at which extreme events
occur (t;u) and the severity (magnitude) of extreme events (t;u) as polynomials in time of
order p and p, respectively while keeping constant. The parameter u is defined as the 90%
quantile of the raw data. This model contains p+p+3 parameters and implies that:
p

(t; u) = u (1 )[1 { (t t 1 ) k } ]{ (t t 1 ) k k }
1

k =0
p

k =0

k =0

(t; u) = { (t t 1 ) k k } { (t t 1 ) k k }(1 )

3.2. TREND DETECTION

(10)

k =0

and

(t)=

Using the models for the parameters analyzed in Section 3.1. estimates for the parameters
, , are produced using MLE. A simple measure, the deviance statistic D, is used to
produce evidence of trends of the previously mentioned categories in the data. The deviance
statistic is defined as: D=2{l1(M1)-l0(M0)}, where l1(M1), l0(M0) are log-likelihoods for
models M1 and M0 respectively (M0 is the likelihood of the model without any trends and M1
the likelihood assuming different trends). If D is lower to f2, where f are the degrees of
freedom of the model with trends incorporated, there is no evidence of the trends examined in
the data. Table 1 shows the results of all the models used. It should be noted that t
corresponds to time (days), while t* to a normalized vector of time considering that on
1/1/1980 all different parameters of Table 1 are stable.
Table 1. Different polynomial temporal trend models for rainfall extremes
Model
Log-likelihood
Time constant
594.816
Linear trend in (=+1t* )
594.816
*
*2
Linear trend in (=+1t +2t )
594.816
Linear trend in (=+1t*)
594.807
*
*
Linear trend in and (=+1t and =+1t )
594.782
Trends in
594.815
(=+1t*+2sin(2t/365)+3cos(2t/365))
Trends in
593.834
(=+1t*+2sin(2t/365)+3cos(2t/365))
Linear trends in (=+1t*) and
594.809
(=(1/)(+1t*))
Trend in number of storms (=+1t*)
594.966
Trend in severity of storms (=+1t*)
594.934
Trends in =+1t* and =+1t*
594.931

0.000
0.000
0.000
0.017
0.068
0.002
1.964
0.014
0.300
0.237
0.229

Estimates of the log-likelihood and the deviance statistic, shown in Table 1, give no
evidence of a polynomial temporal trend of the form described in Section 3.1. in the location
or the scale parameters of the point process model. The largest divergence is observed for the
case where =+1t*+2sin(2t/365)+3cos(2t/365)), but still the deviance statistic is not
considerable. It should be noted that the deviance statistic gives only some evidence of
temporal trend existence and to provide powerful results, different tests have to be used.
4. CONCLUSIONS
In the present study daily rainfall from the city of Thessaloniki, Greece was analyzed
using Extreme Value methodologies to a) detect possible outliers present in the series and b)
investigate the existence of polynomial temporal trends of specified forms in the parameters
of the Extreme Value model used.
Four different methodologies were used to decide upon the inclusion of the largest
observation (the possible outlier) in the data: a) the estimation of xult, b) the conditional
probability of the event estimated through a Bayesian formula, c) a Gumbel plot and d) a q-q
plot for an appropriately selected model for extreme events of the rainfall process.
Finally, simple polynomial temporal trends for the location and the scale parameters were
incorporated in the point process model and a simple test, the deviance statistic D, was used to
identify the significance of such trends.
The main conclusions of the analysis are:
a. Estimates of return levels are significantly lower (up to 20.8% for 1000-year return level),
when the largest observation of 23/11/1985 is omitted from the sample. The distribution of

excesses of the threshold is upper- bounded (the shape parameter is negative) in contrast to
the case including the observation in the analysis.
b. The largest event of 23/11/1985 doesnt seem to be an outlier. All four previously
mentioned approaches (especially b and c) have proven that there is no evidence that this
event should be omitted from the sample.
c. There is no evidence of polynomial temporal trends of the forms described in Section 3.1.
5. ACKNOWLEDGMENTS
The authors want to thank the Dutch National institute RIKZ for providing data of the
Dutch coast. The work described in this publication was supported by the European
Communitys Sixth Framework Programme through the grant to the budget of the Integrated
Project FLOODsite, Contract GOCE-CT-2004-505420. This paper reflects the authors views
and not those of the European Community. Neither the European Community nor any
member of the FLOODsite Consortium is liable for any use of the information in this paper.
6. REFERENCES
Butler, A., 2005, Statistical Modelling of Synthetic Oceanographic Extremes. Ph.D Thesis,
(2005), Lancaster University.
Castillo, E., Hadi, A. S., Balakrishnan, N. and Sarabia J. M., 2005, Extreme Value and
Related Models with Applications in Engineering and Science, Wiley Series in Probability
and Statistics.
Coles, S. G. and Tawn, J. A., 1990, Statistics of coastal flood prevention, Philos. Trans. R.
Soc. Lond., A., 332, pp. 457-476.
Coles, S. G and Tawn, J. A., 1996, A Bayesian analysis of extreme rainfall data, Applied
Statistics, 45, No. 4, pp 463-478.
Coles, S., 2001, An Introduction to Statistical Modeling of Extreme Values, Springer Series in
Statistics.
Davison, A. C. and Smith, R. L., 1990, Models for exceedances over high thresholds (with
discussion), J. R. Stat. Soc., Ser. B, 52, pp. 393-442.
Dixon, M., J. and Tawn, J. A. (1999), The effect of non-stationarity on sea-level estimation,
Applied Statistics, 48, pp. 135-151.
Gaetan, C. and Grigoletto, M., 2004, Smoothing sample extremes with Dynamic Models,
Extremes 7, pp. 221-236
Huerta, G. and Sanso, B., 2005, Time-varying models for extreme values, Department of
Mathematics and Statistics, University of New Mexico, www.ams.ucsc.edu/reports/ams200504.pdf
Pandey, M. D., Van Gelder, P. H. A. J. M. and Vrijling, J. K, 2001, The estimation of
extreme quantiles of wind velocity using L-moments in the peaks-over-threshold approach,
Journal of Structural Safety, 23(2), pp. 179-192.
Reiss, R.-D. and Thomas, M., 2001. Statistical Analysis of Extreme Values with Applications
to Insurance, Finance, Hydrology and Other Fields, Second Edition, Birkhaser.
Robinson, M. E. and Tawn, J. A., 1995, Statistics for exceptional athletic records, Applied
Statistics, 44, pp. 499-511.
Smith, R. L., 2003, Statistics of extremes with applications in Environment, Insurance and
Finance, University of North Carolina, http:/ www.stat.unc.edu/postscript/rs/semstatrls.ps.
Stephenson, A. and Tawn, J., 2004, Bayesian inference for extremes: accounting for the
three Extremal Types, Extremes, 7, No. 4, pp. 291-307.